Sciamlab/landslide-prevention-italy

Pre-Landslide Risk Assessment from Sentinel-2 Imagery

Multi-image vision-language fine-tuning dataset for predicting pre-landslide risk from Sentinel-2 spectral composites. Each sample contains six 512×512 images (one RGB + five spectral indices) paired with a structured JSON assessment of slope hazard, human-impact exposure, and pixel-level areas-of-interest where future failure is most likely.

Generated for fine-tuning LiquidAI/LFM2.5-VL-450M.

---

Dataset Summary

Total samples	111
Train / Test split	89 / 22
Image resolution	512 × 512
Images per sample	6 (RGB + SWIR + NIR + NDMI + NDVI + BSI)
Geography	Italian regions, 2025 landslide events
Source data	Sentinel-2 L2A surface reflectance
Annotator	claude-opus-4-7 (vision)

Risk distribution

Risk level	Count
high	54
low	11
medium	46

---

Dataset Structure

training_data/
├── README.md            ← this file
├── metadata.jsonl       ← one row per sample
├── images/              ← 6 input composites per sample
│   ├── {id}_rgb.png
│   ├── {id}_swir.png
│   ├── {id}_nir.png
│   ├── {id}_ndmi.png
│   ├── {id}_ndvi.png
│   └── {id}_bsi.png
└── aoi/                 ← AOI overlays for visual review (not training input)
    ├── {id}_aoi.png
    ├── {id}_aoi_legend.png
    └── {id}_aoi_combined.png

{id} follows the pattern {event_id}_{town_slug}_{image_date} — e.g. EIT202500057_aymavilles_2025-06-22.

---

Data Fields

Field	Type	Description
id	string	Unique sample identifier
location	string	Original event-folder location string (region > province > town)
town	string	Slugified town name
event_date	string (YYYY-MM-DD)	Date of the recorded landslide event
image_date	string (YYYY-MM-DD)	Sentinel-2 acquisition date
days_before_event	int	Gap between acquisition and event (training-time temporal context)
annotator	string	Producer of the JSON label
split	string	train or test
rgb_path … bsi_path	string	Relative paths to the 6 input images
aoi_path, aoi_legend_path, aoi_combined_path	string	AOI visualisation paths (review only)
spectral_stats	object	Pre-computed NDVI/NDMI/BSI/etc. summary statistics
output	string	JSON-encoded structured assessment (the model's training label)

Output schema

The output string parses to the following object:

{
  "risk_level": "low|medium|high|critical",
  "soil_moisture_anomaly": true,
  "vegetation_stress": false,
  "bare_soil_exposure": true,
  "water_accumulation": false,
  "steep_terrain_visible": true,
  "erosion_evidence": true,
  "clay_soil_detected": true,
  "tct_wetness_anomaly": true,
  "image_quality_limited": false,
  "risk_factors": ["..."],
  "exposure_risk": "low|medium|high|critical",
  "exposed_elements": ["residential area", "main road", "..."],
  "image_size": {"height": 512, "width": 512},
  "areas_of_interest": [
    {"x": 215, "y": 260, "confidence": 0.82, "description": "Bare soil scar on steep slope"}
  ],
  "combined_risk": "low|medium|high|critical",
  "temporal_validity": "high|moderate|low",
  "confidence": 0.78,
  "days_to_event_estimate": "3-10 days",
  "analysis": "2-3 sentence prose summary"
}

The label is intentionally dual-purpose: it captures both the slope-failure hazard and the human-impact exposure, plus pixel-level AOIs in the same 512×512 coordinate frame as the RGB image — so a model can learn to localize hotspots that downstream code can visualize directly.

---

Source Data

• Imagery: Sentinel-2 L2A surface reflectance, Italian regions, 2025.
• Event ground-truth: ISPRA IFFI — Italian National Landslide Inventory.
• Image acquisition: Sentinel-2 scenes selected within ~30 days before each recorded event, prioritising lowest cloud cover.

Annotations

Labels were produced by Claude Opus 4.7 (claude-opus-4-7) via the Anthropic API. Each annotation reads all six composites and a training-time temporal-context note (days_before_event) before emitting the structured JSON above. The model's confidence and temporal_validity fields explicitly account for the staleness gap between acquisition and event.

Considerations for Using

• Spectral indices are proxies, not measurements — NDMI shows surface moisture, not subsurface pore pressure. Predictions describe spectral signatures compatible with failure, not deterministic forecasts.
• No DEM — slope steepness is inferred from RGB shadows.
• Snow/cloud confusion — high-reflectance snow can be misread as bare soil. Filter via the image_quality_limited flag where set.
• AOI coordinates are 1:1 with the 512×512 RGB — overlay them on images/{id}_rgb.png to visualise hotspots.
• Class imbalance — the dataset reflects the geographic distribution of recorded landslide events, not a uniform sample of slopes. Use stratified sampling if training a calibrated probability classifier.

Citation

If you use this dataset, please cite the underlying ISPRA IFFI inventory and the Copernicus Sentinel-2 mission.

---

Generated on 2026-04-29.